Fact · Languages and accents
20languages without training data
Whisper's training data contains no speech for 20 of the 102 languages in the FLEURS benchmark, which caps its language identification accuracy at 80.4%.
Checked against its source on October 17, 2026
In the source's own words
“no training data for 20 of the 102 languages in Fleurs”
Robust Speech Recognition via Large-Scale Weak Supervision (full text), Alec Radford et al., OpenAI (arXiv), 2022, Section 3.6.
Open the original (DOI 10.48550/arXiv.2212.04356)- What it measures
- The FLEURS benchmark covers 102 languages. Whisper was trained on fewer of them, so the benchmark penalizes it for the missing 20.
- What it does not show
- The paper compares Whisper with supervised systems that did have data for those languages, so the 80.4% ceiling is a property of the benchmark setup as well as of Whisper.